Client-local memory → shared context infrastructure
Context is shifting from private memory inside one AI client into a governed layer between agents, IDEs, documents, and tools.
Tectonic shift
Before: Each AI client keeps its own incomplete project memory After: Shared context layer with provenance, governance, and portability Current stage: emerging This New Runtime record is an evidence-linked retrieval unit. Use its canonical page, machine-readable representations, dates, scope, and public source URLs to verify the claim before reusing it.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | unabyss.comsource | primary receipt | source_urls |
| 2 | producthunt.comsource | supporting receipt | source_urls |
| 3 | parallel.aisource | supporting receipt | source_urls |
| 4 | websitemain.claude.comsource | supporting receipt | source_urls |
| 5 | vercel.comsource | supporting receipt | source_urls |
Current state, as of Oct 9, 2026
Where this stands now
strengthening Multiple AI platforms now offer durable, project‑scoped memory layers with provenance and access controls, confirming the shift from client‑local memory to shared context infrastructure.
AI providers such as Anthropic, Grok, Engram, and OpenAI have launched shared, governed project memory layers that persist across sessions and support role‑based sharing. Team Bots and Univer Workspace illustrate concrete products that expose a shared work surface while keeping individual conversations separate. Open‑source and commercial tools like Mem0, Acontext, and SKILL.state provide APIs for durable agent state, observability, and governance.
These developments move context out of isolated client sessions into a durable, observable layer that can be accessed by multiple agents, IDEs, and tools under defined policies. Builders can now rely on a common storage and provenance service for cross‑agent collaboration, reducing manual sync and enabling rollback and audit capabilities.
What changed since the original text
- Anthropic, Grok, Engram, and OpenAI announced durable, project‑scoped memory layers with provenance and governance. 1
- Team Bots package shared workflow context and Univer Workspace provides a shared change surface for people and AI agents. 2
- Acontext offers a separate platform for unified message, file, and skill storage with edit and observability APIs. 3
- Mem0 integrates persistent memory across projects and team members alongside project‑local context tools. 4
- SKILL.state, WikiSkill, Mem0, and other tools describe governed, durable agent state beyond a single response. 5
Open questions
- How will standardized governance models and provenance schemas be adopted across competing platforms?
- What interoperability mechanisms will enable seamless context sharing between heterogeneous AI clients?
Watch next: A major AI vendor releasing a unified, open API spec for shared context layers would solidify the infrastructure shift.
- Weekly update: Multiple AI platforms launch shared, governed project memory layers update, Sep 21, 2026
- Weekly update: Team Bots and Univer Workspace put shared AI work surfaces in focus update, Sep 28, 2026
- Acontext: инфраструктура контекста как отдельный слой для продакшен-агентов archive, Feb 12, 2001
- Add Persistent Memory to Claude Code with Mem0 (5-Minute Setup) archive
- Agent memory становится governed state layer archive, Sep 3, 2001
Rewritten automatically by the New Runtime knowledge desk from this record's weekly updates, releases and archive. The original text below is kept as written (Jul 21, 2026); where the two disagree, this section is newer.
Weekly knowledge desk
Weekly updates
4 reviewed updates to this shift, newest first. Every update is checked against its evidence before it changes the record.
When Claude, Codex, Cursor, and other clients work on the same project, the human often becomes the manual sync layer. This shift captures another option: context lives as standalone infrastructure, and clients connect to it with different permissions.
Evidence
Unabyss shows an early product signal: the context layer can move outside a single client and be connected through an MCP-compatible surface.
Risk
A shared context layer amplifies both usefulness and risk. The GhostWriter piece shows the other side of persistent memory: if memory can be poisoned, it cannot be treated as an ordinary prompt dump.
September 2026 update
What changed
A universal context layer, conversation-shaped projects, and memory observability show context becoming shared infrastructure with an operational lifecycle.
Current boundary
Portability without provenance and access control simply scales a bad handoff. Shared context needs scope, owner, write policy, evidence, and rollback.